intel / intel-xpu-backend-for-triton

OpenAI Triton backend for Intel® GPUs
MIT License
141 stars 43 forks source link

[OCL interface] The OCL interface miss some 2D load variants which is used in Triton #1941

Open chengjunlu opened 2 months ago

chengjunlu commented 2 months ago

New 2D load variants are used in FP8 load. https://github.com/intel/intel-xpu-backend-for-triton/commit/5f088429c8aeb0082c262d812945c2b4e7b86233

New transpose 2D load variants are used in flash attention.

We are using the GENISA if it is not exist in OCL interface for now. Need to summary the new requirements and file JIRA to IGC to track.

vlad-penkin commented 2 months ago

@chengjunlu, @whitneywhtsang could you please file the ticket and share it's number with me offline.

whitneywhtsang commented 2 months ago

@chengjunlu, @whitneywhtsang could you please file the ticket and share it's number with me offline.

Yes, we are still preparing the content for the ticket.

chengjunlu commented 1 month ago

IGC team has confirmed that they are working on the task to support general 2D load interface thru the SPIRV Intel Block IO extension.